C# yolov8 TensorRT +ByteTrack Demo

效果

说明

项目

效果

说明

环境

NVIDIA GeForce RTX 4060 Laptop GPU

cuda12.1+cudnn 8.8.1+TensorRT-8.6.1.6

版本和我不一致的需要重新编译TensorRtExtern.dll，TensorRtExtern源码地址：TensorRT-CSharp-API/src/TensorRtExtern at TensorRtSharp2.0 · guojin-yan/TensorRT-CSharp-API · GitHub

Windows版 CUDA安装参考：Windows版 CUDA安装_win cuda安装-CSDN博客

项目

代码

Form2.cs

using ByteTrack;

using OpenCvSharp;

using System;

using System.Collections.Generic;

using System.Diagnostics;

using System.Drawing;

using System.IO;

using System.Threading;

using System.Windows.Forms;

using TensorRtSharp.Custom;

namespace yolov8_TensorRT_Demo

{

public partial class Form2 : Form

{

public Form2()

{

InitializeComponent();

}

string imgFilter = "*.*|*.bmp;*.jpg;*.jpeg;*.tiff;*.tiff;*.png";

YoloV8 yoloV8;

Mat image;

string image_path = "";

string model_path;

string video_path = "";

string videoFilter = "*.mp4|*.mp4;";

VideoCapture vcapture;

VideoWriter vwriter;

bool saveDetVideo = false;

ByteTracker tracker;

/// <summary>

/// 单图推理

/// </summary>

/// <param name="sender"></param>

/// <param name="e"></param>

private void button2_Click(object sender, EventArgs e)

{

if (image_path == "")

{

return;

}

button2.Enabled = false;

pictureBox2.Image = null;

textBox1.Text = "";

Application.DoEvents();

image = new Mat(image_path);

List<DetectionResult> detResults = yoloV8.Detect(image);

//绘制结果

Mat result_image = image.Clone();

foreach (DetectionResult r in detResults)

{

Cv2.PutText(result_image, $"{r.Class}:{r.Confidence:P0}", new OpenCvSharp.Point(r.Rect.TopLeft.X, r.Rect.TopLeft.Y - 10), HersheyFonts.HersheySimplex, 1, Scalar.Red, 2);

Cv2.Rectangle(result_image, r.Rect, Scalar.Red, thickness: 2);

}

if (pictureBox2.Image != null)

{

pictureBox2.Image.Dispose();

}

pictureBox2.Image = new Bitmap(result_image.ToMemoryStream());

textBox1.Text = yoloV8.DetectTime();

button2.Enabled = true;

}

/// <summary>

/// 窗体加载，初始化

/// </summary>

/// <param name="sender"></param>

/// <param name="e"></param>

private void Form1_Load(object sender, EventArgs e)

{

image_path = "test/dog.jpg";

pictureBox1.Image = new Bitmap(image_path);

model_path = "model/yolov8n.engine";

if (!File.Exists(model_path))

{

//有点耗时，需等待

Nvinfer.OnnxToEngine("model/yolov8n.onnx", 20);

}

yoloV8 = new YoloV8(model_path, "model/lable.txt");

}

/// <summary>

/// 选择图片

/// </summary>

/// <param name="sender"></param>

/// <param name="e"></param>

private void button1_Click_1(object sender, EventArgs e)

{

OpenFileDialog ofd = new OpenFileDialog();

ofd.Filter = imgFilter;

if (ofd.ShowDialog() != DialogResult.OK) return;

pictureBox1.Image = null;

image_path = ofd.FileName;

pictureBox1.Image = new Bitmap(image_path);

textBox1.Text = "";

pictureBox2.Image = null;

}

/// <summary>

/// 选择视频

/// </summary>

/// <param name="sender"></param>

/// <param name="e"></param>

private void button4_Click(object sender, EventArgs e)

{

OpenFileDialog ofd = new OpenFileDialog();

ofd.Filter = videoFilter;

ofd.InitialDirectory = Application.StartupPath + "\\test";

if (ofd.ShowDialog() != DialogResult.OK) return;

video_path = ofd.FileName;

textBox1.Text = "";

pictureBox1.Image = null;

pictureBox2.Image = null;

button3_Click(null, null);

}

/// <summary>

/// 视频推理

/// </summary>

/// <param name="sender"></param>

/// <param name="e"></param>

private void button3_Click(object sender, EventArgs e)

{

if (video_path == "")

{

return;

}

textBox1.Text = "开始检测";

Application.DoEvents();

Thread thread = new Thread(new ThreadStart(VideoDetection));

thread.Start();

thread.Join();

textBox1.Text = "检测完成！";

}

void VideoDetection()

{

vcapture = new VideoCapture(video_path);

if (!vcapture.IsOpened())

{

MessageBox.Show("打开视频文件失败");

return;

}

tracker = new ByteTracker((int)vcapture.Fps, 200);

Mat frame = new Mat();

List<DetectionResult> detResults;

// 获取视频的fps

double videoFps = vcapture.Get(VideoCaptureProperties.Fps);

// 计算等待时间（毫秒）

int delay = (int)(1000 / videoFps);

Stopwatch _stopwatch = new Stopwatch();

if (checkBox1.Checked)

{

vwriter = new VideoWriter("out.mp4", FourCC.X264, vcapture.Fps, new OpenCvSharp.Size(vcapture.FrameWidth, vcapture.FrameHeight));

saveDetVideo = true;

}

else

{

saveDetVideo = false;

}

while (vcapture.Read(frame))

{

if (frame.Empty())

{

MessageBox.Show("读取失败");

return;

}

_stopwatch.Restart();

delay = (int)(1000 / videoFps);

detResults = yoloV8.Detect(frame);

//绘制结果

//foreach (DetectionResult r in detResults)

//{

// Cv2.PutText(frame, $"{r.Class}:{r.Confidence:P0}", new OpenCvSharp.Point(r.Rect.TopLeft.X, r.Rect.TopLeft.Y - 10), HersheyFonts.HersheySimplex, 1, Scalar.Red, 2);

// Cv2.Rectangle(frame, r.Rect, Scalar.Red, thickness: 2);

//}

Cv2.PutText(frame, "preprocessTime:" + yoloV8.preprocessTime.ToString("F2") + "ms", new OpenCvSharp.Point(10, 30), HersheyFonts.HersheySimplex, 1, Scalar.Red, 2);

Cv2.PutText(frame, "inferTime:" + yoloV8.inferTime.ToString("F2") + "ms", new OpenCvSharp.Point(10, 70), HersheyFonts.HersheySimplex, 1, Scalar.Red, 2);

Cv2.PutText(frame, "postprocessTime:" + yoloV8.postprocessTime.ToString("F2") + "ms", new OpenCvSharp.Point(10, 110), HersheyFonts.HersheySimplex, 1, Scalar.Red, 2);

Cv2.PutText(frame, "totalTime:" + yoloV8.totalTime.ToString("F2") + "ms", new OpenCvSharp.Point(10, 150), HersheyFonts.HersheySimplex, 1, Scalar.Red, 2);

Cv2.PutText(frame, "video fps:" + videoFps.ToString("F2"), new OpenCvSharp.Point(10, 190), HersheyFonts.HersheySimplex, 1, Scalar.Red, 2);

Cv2.PutText(frame, "det fps:" + yoloV8.detFps.ToString("F2"), new OpenCvSharp.Point(10, 230), HersheyFonts.HersheySimplex, 1, Scalar.Red, 2);

List<Track> track = new List<Track>();

Track temp;

foreach (DetectionResult r in detResults)

{

RectBox _box = new RectBox(r.Rect.X, r.Rect.Y, r.Rect.Width, r.Rect.Height);

temp = new Track(_box, r.Confidence, ("label", r.ClassId), ("name", r.Class));

track.Add(temp);

}

var trackOutputs = tracker.Update(track);

foreach (var t in trackOutputs)

{

Rect rect = new Rect((int)t.RectBox.X, (int)t.RectBox.Y, (int)t.RectBox.Width, (int)t.RectBox.Height);

//string txt = $"{t["name"]}-{t.TrackId}:{t.Score:P0}";

string txt = $"{t["name"]}-{t.TrackId}";

Cv2.PutText(frame, txt, new OpenCvSharp.Point(rect.TopLeft.X, rect.TopLeft.Y - 10), HersheyFonts.HersheySimplex, 1, Scalar.Red, 2);

Cv2.Rectangle(frame, rect, Scalar.Red, thickness: 2);

}

if (saveDetVideo)

{

vwriter.Write(frame);

}

Cv2.ImShow("DetectionResult", frame);

// for test

// delay = 1;

delay = (int)(delay - _stopwatch.ElapsedMilliseconds);

if (delay <= 0)

{

delay = 1;

}

//Console.WriteLine("delay:" + delay.ToString()) ;

if (Cv2.WaitKey(delay) == 27)

{

break; // 如果按下ESC，退出循环

}

Cv2.DestroyAllWindows();

vcapture.Release();

if (saveDetVideo)

{

vwriter.Release();

}

using ByteTrack;
using OpenCvSharp;
using System;
using System.Collections.Generic;
using System.Diagnostics;
using System.Drawing;
using System.IO;
using System.Threading;
using System.Windows.Forms;
using TensorRtSharp.Custom;

namespace yolov8_TensorRT_Demo
{
    public partial class Form2 : Form
    {
        public Form2()
        {
            InitializeComponent();
        }

        string imgFilter = "*.*|*.bmp;*.jpg;*.jpeg;*.tiff;*.tiff;*.png";

        YoloV8 yoloV8;
        Mat image;

        string image_path = "";
        string model_path;

        string video_path = "";
        string videoFilter = "*.mp4|*.mp4;";
        VideoCapture vcapture;
        VideoWriter vwriter;
        bool saveDetVideo = false;
        ByteTracker tracker;


        /// <summary>
        /// 单图推理
        /// </summary>
        /// <param name="sender"></param>
        /// <param name="e"></param>
        private void button2_Click(object sender, EventArgs e)
        {

            if (image_path == "")
            {
                return;
            }

            button2.Enabled = false;
            pictureBox2.Image = null;
            textBox1.Text = "";

            Application.DoEvents();

            image = new Mat(image_path);

            List<DetectionResult> detResults = yoloV8.Detect(image);

            //绘制结果
            Mat result_image = image.Clone();
            foreach (DetectionResult r in detResults)
            {
                Cv2.PutText(result_image, $"{r.Class}:{r.Confidence:P0}", new OpenCvSharp.Point(r.Rect.TopLeft.X, r.Rect.TopLeft.Y - 10), HersheyFonts.HersheySimplex, 1, Scalar.Red, 2);
                Cv2.Rectangle(result_image, r.Rect, Scalar.Red, thickness: 2);
            }

            if (pictureBox2.Image != null)
            {
                pictureBox2.Image.Dispose();
            }
            pictureBox2.Image = new Bitmap(result_image.ToMemoryStream());
            textBox1.Text = yoloV8.DetectTime();

            button2.Enabled = true;

        }

        /// <summary>
        /// 窗体加载，初始化
        /// </summary>
        /// <param name="sender"></param>
        /// <param name="e"></param>
        private void Form1_Load(object sender, EventArgs e)
        {
            image_path = "test/dog.jpg";
            pictureBox1.Image = new Bitmap(image_path);

            model_path = "model/yolov8n.engine";

            if (!File.Exists(model_path))
            {
                //有点耗时，需等待
                Nvinfer.OnnxToEngine("model/yolov8n.onnx", 20);
            }

            yoloV8 = new YoloV8(model_path, "model/lable.txt");

        }

        /// <summary>
        /// 选择图片
        /// </summary>
        /// <param name="sender"></param>
        /// <param name="e"></param>
        private void button1_Click_1(object sender, EventArgs e)
        {
            OpenFileDialog ofd = new OpenFileDialog();
            ofd.Filter = imgFilter;
            if (ofd.ShowDialog() != DialogResult.OK) return;

            pictureBox1.Image = null;

            image_path = ofd.FileName;
            pictureBox1.Image = new Bitmap(image_path);

            textBox1.Text = "";
            pictureBox2.Image = null;
        }

        /// <summary>
        /// 选择视频
        /// </summary>
        /// <param name="sender"></param>
        /// <param name="e"></param>
        private void button4_Click(object sender, EventArgs e)
        {
            OpenFileDialog ofd = new OpenFileDialog();
            ofd.Filter = videoFilter;
            ofd.InitialDirectory = Application.StartupPath + "\\test";
            if (ofd.ShowDialog() != DialogResult.OK) return;

            video_path = ofd.FileName;

            textBox1.Text = "";
            pictureBox1.Image = null;
            pictureBox2.Image = null;

            button3_Click(null, null);

        }

        /// <summary>
        /// 视频推理
        /// </summary>
        /// <param name="sender"></param>
        /// <param name="e"></param>
        private void button3_Click(object sender, EventArgs e)
        {
            if (video_path == "")
            {
                return;
            }

            textBox1.Text = "开始检测";

            Application.DoEvents();

            Thread thread = new Thread(new ThreadStart(VideoDetection));

            thread.Start();
            thread.Join();

            textBox1.Text = "检测完成！";
        }

        void VideoDetection()
        {
            vcapture = new VideoCapture(video_path);
            if (!vcapture.IsOpened())
            {
                MessageBox.Show("打开视频文件失败");
                return;
            }

            tracker = new ByteTracker((int)vcapture.Fps, 200);

            Mat frame = new Mat();
            List<DetectionResult> detResults;

            // 获取视频的fps
            double videoFps = vcapture.Get(VideoCaptureProperties.Fps);
            // 计算等待时间（毫秒）
            int delay = (int)(1000 / videoFps);
            Stopwatch _stopwatch = new Stopwatch();

            if (checkBox1.Checked)
            {
                vwriter = new VideoWriter("out.mp4", FourCC.X264, vcapture.Fps, new OpenCvSharp.Size(vcapture.FrameWidth, vcapture.FrameHeight));
                saveDetVideo = true;
            }
            else
            {
                saveDetVideo = false;
            }

            while (vcapture.Read(frame))
            {
                if (frame.Empty())
                {
                    MessageBox.Show("读取失败");
                    return;
                }

                _stopwatch.Restart();

                delay = (int)(1000 / videoFps);

                detResults = yoloV8.Detect(frame);

                //绘制结果
                //foreach (DetectionResult r in detResults)
                //{
                //    Cv2.PutText(frame, $"{r.Class}:{r.Confidence:P0}", new OpenCvSharp.Point(r.Rect.TopLeft.X, r.Rect.TopLeft.Y - 10), HersheyFonts.HersheySimplex, 1, Scalar.Red, 2);
                //    Cv2.Rectangle(frame, r.Rect, Scalar.Red, thickness: 2);
                //}

                Cv2.PutText(frame, "preprocessTime:" + yoloV8.preprocessTime.ToString("F2") + "ms", new OpenCvSharp.Point(10, 30), HersheyFonts.HersheySimplex, 1, Scalar.Red, 2);
                Cv2.PutText(frame, "inferTime:" + yoloV8.inferTime.ToString("F2") + "ms", new OpenCvSharp.Point(10, 70), HersheyFonts.HersheySimplex, 1, Scalar.Red, 2);
                Cv2.PutText(frame, "postprocessTime:" + yoloV8.postprocessTime.ToString("F2") + "ms", new OpenCvSharp.Point(10, 110), HersheyFonts.HersheySimplex, 1, Scalar.Red, 2);
                Cv2.PutText(frame, "totalTime:" + yoloV8.totalTime.ToString("F2") + "ms", new OpenCvSharp.Point(10, 150), HersheyFonts.HersheySimplex, 1, Scalar.Red, 2);
                Cv2.PutText(frame, "video fps:" + videoFps.ToString("F2"), new OpenCvSharp.Point(10, 190), HersheyFonts.HersheySimplex, 1, Scalar.Red, 2);
                Cv2.PutText(frame, "det fps:" + yoloV8.detFps.ToString("F2"), new OpenCvSharp.Point(10, 230), HersheyFonts.HersheySimplex, 1, Scalar.Red, 2);

                List<Track> track = new List<Track>();
                Track temp;
                foreach (DetectionResult r in detResults)
                {
                    RectBox _box = new RectBox(r.Rect.X, r.Rect.Y, r.Rect.Width, r.Rect.Height);
                    temp = new Track(_box, r.Confidence, ("label", r.ClassId), ("name", r.Class));
                    track.Add(temp);
                }

                var trackOutputs = tracker.Update(track);

                foreach (var t in trackOutputs)
                {
                    Rect rect = new Rect((int)t.RectBox.X, (int)t.RectBox.Y, (int)t.RectBox.Width, (int)t.RectBox.Height);
                    //string txt = $"{t["name"]}-{t.TrackId}:{t.Score:P0}";
                    string txt = $"{t["name"]}-{t.TrackId}";
                    Cv2.PutText(frame, txt, new OpenCvSharp.Point(rect.TopLeft.X, rect.TopLeft.Y - 10), HersheyFonts.HersheySimplex, 1, Scalar.Red, 2);
                    Cv2.Rectangle(frame, rect, Scalar.Red, thickness: 2);
                }

                if (saveDetVideo)
                {
                    vwriter.Write(frame);
                }

                Cv2.ImShow("DetectionResult", frame);

                // for test
                // delay = 1;
                delay = (int)(delay - _stopwatch.ElapsedMilliseconds);
                if (delay <= 0)
                {
                    delay = 1;
                }
                //Console.WriteLine("delay:" + delay.ToString()) ;
                if (Cv2.WaitKey(delay) == 27)
                {
                    break; // 如果按下ESC，退出循环
                }
            }

            Cv2.DestroyAllWindows();
            vcapture.Release();
            if (saveDetVideo)
            {
                vwriter.Release();
            }

        }

    }

}

YoloV8.cs

using OpenCvSharp;
using OpenCvSharp.Dnn;
using System;
using System.Collections.Generic;
using System.Diagnostics;
using System.IO;
using System.Linq;
using System.Text;
using TensorRtSharp.Custom;

namespace yolov8_TensorRT_Demo
{
    public class YoloV8
    {

        float[] input_tensor_data;
        float[] outputData;
        List<DetectionResult> detectionResults;

        int input_height;
        int input_width;

        Nvinfer predictor;

        public string[] class_names;
        int class_num;
        int box_num;

        float conf_threshold;
        float nms_threshold;

        float ratio_height;
        float ratio_width;

        public double preprocessTime;
        public double inferTime;
        public double postprocessTime;
        public double totalTime;
        public double detFps;

        public String DetectTime()
        {
            StringBuilder stringBuilder = new StringBuilder();
            stringBuilder.AppendLine($"Preprocess: {preprocessTime:F2}ms");
            stringBuilder.AppendLine($"Infer: {inferTime:F2}ms");
            stringBuilder.AppendLine($"Postprocess: {postprocessTime:F2}ms");
            stringBuilder.AppendLine($"Total: {totalTime:F2}ms");

            return stringBuilder.ToString();
        }

        public YoloV8(string model_path, string classer_path)
        {
            predictor = new Nvinfer(model_path);

            class_names = File.ReadAllLines(classer_path, Encoding.UTF8);
            class_num = class_names.Length;

            input_height = 640;
            input_width = 640;

            box_num = 8400;

            conf_threshold = 0.25f;
            nms_threshold = 0.5f;

            detectionResults = new List<DetectionResult>();
        }

        void Preprocess(Mat image)
        {
            //图片缩放
            int height = image.Rows;
            int width = image.Cols;
            Mat temp_image = image.Clone();
            if (height > input_height || width > input_width)
            {
                float scale = Math.Min((float)input_height / height, (float)input_width / width);
                OpenCvSharp.Size new_size = new OpenCvSharp.Size((int)(width * scale), (int)(height * scale));
                Cv2.Resize(image, temp_image, new_size);
            }
            ratio_height = (float)height / temp_image.Rows;
            ratio_width = (float)width / temp_image.Cols;
            Mat input_img = new Mat();
            Cv2.CopyMakeBorder(temp_image, input_img, 0, input_height - temp_image.Rows, 0, input_width - temp_image.Cols, BorderTypes.Constant, 0);

            //归一化
            input_img.ConvertTo(input_img, MatType.CV_32FC3, 1.0 / 255);

            input_tensor_data = Common.ExtractMat(input_img);

            input_img.Dispose();
            temp_image.Dispose();
        }

        void Postprocess(float[] outputData)
        {
            detectionResults.Clear();

            float[] data = Common.Transpose(outputData, class_num + 4, box_num);

            float[] confidenceInfo = new float[class_num];
            float[] rectData = new float[4];

            List<DetectionResult> detResults = new List<DetectionResult>();

            for (int i = 0; i < box_num; i++)
            {
                Array.Copy(data, i * (class_num + 4), rectData, 0, 4);
                Array.Copy(data, i * (class_num + 4) + 4, confidenceInfo, 0, class_num);

                float score = confidenceInfo.Max(); // 获取最大值

                int maxIndex = Array.IndexOf(confidenceInfo, score); // 获取最大值的位置

                int _centerX = (int)(rectData[0] * ratio_width);
                int _centerY = (int)(rectData[1] * ratio_height);
                int _width = (int)(rectData[2] * ratio_width);
                int _height = (int)(rectData[3] * ratio_height);

                detResults.Add(new DetectionResult(
                   maxIndex,
                   class_names[maxIndex],
                   new Rect(_centerX - _width / 2, _centerY - _height / 2, _width, _height),
                   score));
            }

            //NMS
            CvDnn.NMSBoxes(detResults.Select(x => x.Rect), detResults.Select(x => x.Confidence), conf_threshold, nms_threshold, out int[] indices);
            detResults = detResults.Where((x, index) => indices.Contains(index)).ToList();

            detectionResults = detResults;
        }

        internal List<DetectionResult> Detect(Mat image)
        {

            var t1 = Cv2.GetTickCount();

            Stopwatch stopwatch = new Stopwatch();
            stopwatch.Start();

            Preprocess(image);

            preprocessTime = stopwatch.Elapsed.TotalMilliseconds;
            stopwatch.Restart();

            predictor.LoadInferenceData("images", input_tensor_data);

            predictor.infer();

            inferTime = stopwatch.Elapsed.TotalMilliseconds;
            stopwatch.Restart();

            outputData = predictor.GetInferenceResult("output0");

            Postprocess(outputData);

            postprocessTime = stopwatch.Elapsed.TotalMilliseconds;
            stopwatch.Stop();

            totalTime = preprocessTime + inferTime + postprocessTime;

            detFps = (double)stopwatch.Elapsed.TotalSeconds / (double)stopwatch.Elapsed.Ticks;

            var t2 = Cv2.GetTickCount();

            detFps = 1 / ((t2 - t1) / Cv2.GetTickFrequency());

            return detectionResults;

        }

    }
}

ByteTracker.cs

using System;
using System.Collections.Generic;
using System.Linq;
using System.Text;
using System.Threading.Tasks;

namespace ByteTrack
{
    public class ByteTracker
    {
        readonly float _trackThresh;
        readonly float _highThresh;
        readonly float _matchThresh;
        readonly int _maxTimeLost;

        int _frameId = 0;
        int _trackIdCount = 0;

        readonly List<Track> _trackedTracks = new List<Track>(100);
        readonly List<Track> _lostTracks = new List<Track>(100);
        List<Track> _removedTracks = new List<Track>(100);

        public ByteTracker(int frameRate = 30, int trackBuffer = 30, float trackThresh = 0.5f, float highThresh = 0.6f, float matchThresh = 0.8f)
        {
            _trackThresh = trackThresh;
            _highThresh = highThresh;
            _matchThresh = matchThresh;
            _maxTimeLost = (int)(frameRate / 30.0 * trackBuffer);
        }

        /// <summary>
        /// 
        /// </summary>
        /// <param name="objects"></param>
        /// <returns></returns>
        public IList<Track> Update(List<Track> tracks)
        {
            #region Step 1: Get detections 
            _frameId++;

            // Create new Tracks using the result of object detection
            List<Track> detTracks = new List<Track>();
            List<Track> detLowTracks = new List<Track>();

            foreach (var obj in tracks)
            {
                if (obj.Score >= _trackThresh)
                {
                    detTracks.Add(obj);
                }
                else
                {
                    detLowTracks.Add(obj);
                }
            }

            // Create lists of existing STrack
            List<Track> activeTracks = new List<Track>();
            List<Track> nonActiveTracks = new List<Track>();

            foreach (var trackedTrack in _trackedTracks)
            {
                if (!trackedTrack.IsActivated)
                {
                    nonActiveTracks.Add(trackedTrack);
                }
                else
                {
                    activeTracks.Add(trackedTrack);
                }
            }

            var trackPool = activeTracks.Union(_lostTracks).ToArray();

            // Predict current pose by KF
            foreach (var track in trackPool)
            {
                track.Predict();
            }
            #endregion

            #region Step 2: First association, with IoU 
            List<Track> currentTrackedTracks = new List<Track>();
            Track[] remainTrackedTracks;
            Track[] remainDetTracks;
            List<Track> refindTracks = new List<Track>();
            {
                var dists = CalcIouDistance(trackPool, detTracks);
                LinearAssignment(dists, trackPool.Length, detTracks.Count, _matchThresh,
                    out var matchesIdx,
                    out var unmatchTrackIdx,
                    out var unmatchDetectionIdx);

                foreach (var matchIdx in matchesIdx)
                {
                    var track = trackPool[matchIdx[0]];
                    var det = detTracks[matchIdx[1]];
                    if (track.State == TrackState.Tracked)
                    {
                        track.Update(det, _frameId);
                        currentTrackedTracks.Add(track);
                    }
                    else
                    {
                        track.ReActivate(det, _frameId);
                        refindTracks.Add(track);
                    }
                }

                remainDetTracks = unmatchDetectionIdx.Select(unmatchIdx => detTracks[unmatchIdx]).ToArray();
                remainTrackedTracks = unmatchTrackIdx
                    .Where(unmatchIdx => trackPool[unmatchIdx].State == TrackState.Tracked)
                    .Select(unmatchIdx => trackPool[unmatchIdx])
                    .ToArray();
            }
            #endregion

            #region Step 3: Second association, using low score dets 
            List<Track> currentLostTracks = new List<Track>();
            {
                var dists = CalcIouDistance(remainTrackedTracks, detLowTracks);
                LinearAssignment(dists, remainTrackedTracks.Length, detLowTracks.Count, 0.5f,
                                 out var matchesIdx,
                                 out var unmatchTrackIdx,
                                 out var unmatchDetectionIdx);

                foreach (var matchIdx in matchesIdx)
                {
                    var track = remainTrackedTracks[matchIdx[0]];
                    var det = detLowTracks[matchIdx[1]];
                    if (track.State == TrackState.Tracked)
                    {
                        track.Update(det, _frameId);
                        currentTrackedTracks.Add(track);
                    }
                    else
                    {
                        track.ReActivate(det, _frameId);
                        refindTracks.Add(track);
                    }
                }

                foreach (var unmatchTrack in unmatchTrackIdx)
                {
                    var track = remainTrackedTracks[unmatchTrack];
                    if (track.State != TrackState.Lost)
                    {
                        track.MarkAsLost();
                        currentLostTracks.Add(track);
                    }
                }
            }
            #endregion

            #region Step 4: Init new tracks 
            List<Track> currentRemovedTracks = new List<Track>();
            {
                // Deal with unconfirmed tracks, usually tracks with only one beginning frame
                var dists = CalcIouDistance(nonActiveTracks, remainDetTracks);
                LinearAssignment(dists, nonActiveTracks.Count, remainDetTracks.Length, 0.7f,
                                 out var matchesIdx,
                                 out var unmatchUnconfirmedIdx,
                                 out var unmatchDetectionIdx);

                foreach (var matchIdx in matchesIdx)
                {
                    nonActiveTracks[matchIdx[0]].Update(remainDetTracks[matchIdx[1]], _frameId);
                    currentTrackedTracks.Add(nonActiveTracks[matchIdx[0]]);
                }

                foreach (var unmatchIdx in unmatchUnconfirmedIdx)
                {
                    var track = nonActiveTracks[unmatchIdx];
                    track.MarkAsRemoved();
                    currentRemovedTracks.Add(track);
                }

                // Add new stracks
                foreach (var unmatchIdx in unmatchDetectionIdx)
                {
                    var track = remainDetTracks[unmatchIdx];
                    if (track.Score < _highThresh)
                        continue;

                    _trackIdCount++;
                    track.Activate(_frameId, _trackIdCount);
                    currentTrackedTracks.Add(track);
                }
            }
            #endregion

            #region Step 5: Update state
            foreach (var lostTrack in _lostTracks)
            {
                if (_frameId - lostTrack.FrameId > _maxTimeLost)
                {
                    lostTrack.MarkAsRemoved();
                    currentRemovedTracks.Add(lostTrack);
                }
            }

            var trackedTracks = currentTrackedTracks.Union(refindTracks).ToArray();
            var lostTracks = _lostTracks.Except(trackedTracks).Union(currentLostTracks).Except(_removedTracks).ToArray();
            _removedTracks = _removedTracks.Union(currentRemovedTracks).ToList();
            RemoveDuplicateStracks(trackedTracks, lostTracks);
            #endregion

            return _trackedTracks.Where(track => track.IsActivated).ToArray();
        }

        /// <summary>
        /// 
        /// </summary>
        /// <param name="aTracks"></param>
        /// <param name="bTracks"></param>
        /// <param name="aResults"></param>
        /// <param name="bResults"></param>
        void RemoveDuplicateStracks(IList<Track> aTracks, IList<Track> bTracks)
        {
            _trackedTracks.Clear();
            _lostTracks.Clear();

            List<(int, int)> overlappingCombinations;
            var ious = CalcIouDistance(aTracks, bTracks);

            if (ious is null)
                overlappingCombinations = new List<(int, int)>();
            else
            {
                var rows = ious.GetLength(0);
                var cols = ious.GetLength(1);
                overlappingCombinations = new List<(int, int)>(rows * cols / 2);
                for (var i = 0; i < rows; i++)
                    for (var j = 0; j < cols; j++)
                        if (ious[i, j] < 0.15f)
                            overlappingCombinations.Add((i, j));
            }

            var aOverlapping = aTracks.Select(x => false).ToArray();
            var bOverlapping = bTracks.Select(x => false).ToArray();

            foreach (var (aIdx, bIdx) in overlappingCombinations)
            {
                var timep = aTracks[aIdx].FrameId - aTracks[aIdx].StartFrameId;
                var timeq = bTracks[bIdx].FrameId - bTracks[bIdx].StartFrameId;
                if (timep > timeq)
                    bOverlapping[bIdx] = true;
                else
                    aOverlapping[aIdx] = true;
            }

            for (var ai = 0; ai < aTracks.Count; ai++)
                if (!aOverlapping[ai])
                    _trackedTracks.Add(aTracks[ai]);

            for (var bi = 0; bi < bTracks.Count; bi++)
                if (!bOverlapping[bi])
                    _lostTracks.Add(bTracks[bi]);
        }

        /// <summary>
        /// 
        /// </summary>
        /// <param name="costMatrix"></param>
        /// <param name="costMatrixSize"></param>
        /// <param name="costMatrixSizeSize"></param>
        /// <param name="thresh"></param>
        /// <param name="matches"></param>
        /// <param name="aUnmatched"></param>
        /// <param name="bUnmatched"></param>
        void LinearAssignment(float[,] costMatrix, int costMatrixSize, int costMatrixSizeSize, float thresh, out IList<int[]> matches, out IList<int> aUnmatched, out IList<int> bUnmatched)
        {
            matches = new List<int[]>();
            if (costMatrix is null)
            {
                aUnmatched = Enumerable.Range(0, costMatrixSize).ToArray();
                bUnmatched = Enumerable.Range(0, costMatrixSizeSize).ToArray();
                return;
            }

            bUnmatched = new List<int>();
            aUnmatched = new List<int>();

            var (rowsol, colsol) = Lapjv.Exec(costMatrix, true, thresh);

            for (var i = 0; i < rowsol.Length; i++)
            {
                if (rowsol[i] >= 0)
                    matches.Add(new int[] { i, rowsol[i] });
                else
                    aUnmatched.Add(i);
            }

            for (var i = 0; i < colsol.Length; i++)
                if (colsol[i] < 0)
                    bUnmatched.Add(i);
        }

        /// <summary>
        /// 
        /// </summary>
        /// <param name="aRects"></param>
        /// <param name="bRects"></param>
        /// <returns></returns>
        static float[,] CalcIous(IList<RectBox> aRects, IList<RectBox> bRects)
        {
            if (aRects.Count * bRects.Count == 0) return null;

            var ious = new float[aRects.Count, bRects.Count];
            for (var bi = 0; bi < bRects.Count; bi++)
                for (var ai = 0; ai < aRects.Count; ai++)
                    ious[ai, bi] = bRects[bi].CalcIoU(aRects[ai]);

            return ious;
        }

        /// <summary>
        /// 
        /// </summary>
        /// <param name="aTtracks"></param>
        /// <param name="bTracks"></param>
        /// <returns></returns>
        static float[,] CalcIouDistance(IEnumerable<Track> aTtracks, IEnumerable<Track> bTracks)
        {
            var aRects = aTtracks.Select(x => x.RectBox).ToArray();
            var bRects = bTracks.Select(x => x.RectBox).ToArray();

            var ious = CalcIous(aRects, bRects);
            if (ious is null) return null;

            var rows = ious.GetLength(0);
            var cols = ious.GetLength(1);
            var matrix = new float[rows, cols];
            for (var i = 0; i < rows; i++)
                for (var j = 0; j < cols; j++)
                    matrix[i, j] = 1 - ious[i, j];

            return matrix;
        }
    }
}

下载

源码下载

参考

https://github.com/devhxj/Yolo8-ByteTrack-CSharp

https://github.com/guojin-yan/TensorRT-CSharp-API